x86: Support RDTSCP for benchtests

Message ID CAMe9rOoUHBsr9R048eNGuCWmbuu_Z9AwPkStMs06Mfr99ki_OQ@mail.gmail.com
State New
Headers show
Series
  • x86: Support RDTSCP for benchtests
Related show

Commit Message

H.J. Lu Oct. 23, 2018, 9:29 p.m.
On 10/23/18, Szabolcs Nagy <Szabolcs.Nagy@arm.com> wrote:
> On 23/10/18 11:58, H.J. Lu wrote:

>> On 10/23/18, Siddhesh Poyarekar <siddhesh@gotplt.org> wrote:

>>> On 23/10/18 2:34 PM, Florian Weimer wrote:

>>>> Shouldn't the benchtests use clock_gettime anyway, to avoid issues in

>>>> case the TSC is not synchronized across cores?

>>>

>>> There's an option USE_CLOCK_GETTIME to make benchtests do that, but

>>> otherwise it uses the hp_timing bits by default.

>>

>> I want something better that rdtsc and very low overhead since some bench

>> tests only last a few cycles.  Adding lfence may make timing data look

>> like

>> noise.

>>

>

> ideally bench test should be fixed so clock_gettime gives

> stable enough results.

>

> target specific timers are not always available and their

> results are hard to interpret compared to a standard api

> that returns wall clock time in sane units.

>


Here is a simple patch to support RDTSCP for benchtests.
OK for trunk?

-- 
H.J.

Comments

Siddhesh Poyarekar Oct. 24, 2018, 8:29 a.m. | #1
On 24/10/18 2:59 AM, H.J. Lu wrote:
> Here is a simple patch to support RDTSCP for benchtests.

> OK for trunk?


Looks fine.

Siddhesh

Patch

From a1cf1cb1c86cfc99b81ce6e1caf5807d2ec25c08 Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Mon, 22 Oct 2018 01:13:38 -0700
Subject: [PATCH] x86: Support RDTSCP for benchtests

RDTSCP waits until all previous instructions have executed and all
previous loads are globally visible before reading the counter.  RDTSC
doesn't wait until all previous instructions have been executed before
reading the counter.  All x86 processors since 2010 support RDTSCP
instruction.  This patch adds RDTSCP support to benchtests.

	* benchtests/Makefile (CPPFLAGS-nonlib): Add -DUSE_RDTSCP if
	USE_RDTSCP is defined.
	* sysdeps/x86/hp-timing.h (HP_TIMING_NOW): Use RDTSCP if
	USE_RDTSCP is defined.
---
 benchtests/Makefile     |  6 ++++++
 benchtests/README       |  9 +++++++++
 sysdeps/x86/hp-timing.h | 14 +++++++++++++-
 3 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/benchtests/Makefile b/benchtests/Makefile
index bcd6a9c26d..45aeb5febe 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -131,6 +131,12 @@  CPPFLAGS-nonlib += -DDURATION=$(BENCH_DURATION) -D_ISOMAC
 # HP_TIMING if it is available.
 ifdef USE_CLOCK_GETTIME
 CPPFLAGS-nonlib += -DUSE_CLOCK_GETTIME
+else
+# On x86 processors, use RDTSCP, instead of RDTSC, to measure performance
+# of functions.  All x86 processors since 2010 support RDTSCP instruction.
+ifdef USE_RDTSCP
+CPPFLAGS-nonlib += -DUSE_RDTSCP
+endif
 endif
 
 DETAILED_OPT :=
diff --git a/benchtests/README b/benchtests/README
index 4ddff794d1..aaf0b659e2 100644
--- a/benchtests/README
+++ b/benchtests/README
@@ -34,6 +34,15 @@  the benchmark to use clock_gettime by invoking make as follows:
 
 Again, one must run `make bench-clean' before changing the measurement method.
 
+On x86 processors, RDTSCP instruction provides more precise timing data
+than RDTSC instruction.  All x86 processors since 2010 support RDTSCP
+instruction.  One can force the benchmark to use RDTSCP by invoking make
+as follows:
+
+  $ make USE_RDTSCP=1 bench
+
+One must run `make bench-clean' before changing the measurement method.
+
 Running benchmarks on another target:
 ====================================
 
diff --git a/sysdeps/x86/hp-timing.h b/sysdeps/x86/hp-timing.h
index 77a1360748..0aa6f5e3f8 100644
--- a/sysdeps/x86/hp-timing.h
+++ b/sysdeps/x86/hp-timing.h
@@ -40,7 +40,19 @@  typedef unsigned long long int hp_timing_t;
 
    NB: Use __builtin_ia32_rdtsc directly since including <x86intrin.h>
    makes building glibc very slow.  */
-# define HP_TIMING_NOW(Var)	((Var) = __builtin_ia32_rdtsc ())
+# ifdef USE_RDTSCP
+/* RDTSCP waits until all previous instructions have executed and all
+   previous loads are globally visible before reading the counter.
+   RDTSC doesn't wait until all previous instructions have been executed
+   before reading the counter.  */
+#  define HP_TIMING_NOW(Var) \
+  (__extension__ ({				\
+    unsigned int __aux;				\
+    (Var) = __builtin_ia32_rdtscp (&__aux);	\
+  }))
+# else
+#  define HP_TIMING_NOW(Var) ((Var) = __builtin_ia32_rdtsc ())
+# endif
 
 # include <hp-timing-common.h>
 #else
-- 
2.17.2